A German fine-tune of Kokoro-82M, exported to ONNX for use with Crane, a high-performance LLM and multimodal inference framework. Kokoro's upstream weights and the misaki G2P frontend that normally drives them have no German support, so this repository pairs a German-specific fine-tune of the model with Crane's own hand-built German G2P (lexicon + rules, no espeak-ng dependency) to produce compatible input.

This model is German-only. It ships a single German voice (df_kerstin), and the fine-tune updates the full backbone, including the text encoder, on German audio and German phonemes. Whether any of the upstream model's English capability survives this process hasn't been tested; treat this checkpoint as German-only rather than bilingual.

This is not a drop-in replacement for the upstream Kokoro-82M-v1.0-ONNX checkpoint. It will not produce sensible German (or English) output through the stock English-trained weights, and it was not trained by the upstream Kokoro/misaki authors.

The German fine-tune was produced with kikiri-tts (formerly kokoro-deutsch), a reproducible two-stage StyleTTS2 fine-tuning pipeline for adapting Kokoro-82M to a new language. Unlike a frozen-backbone style/speaker embedding, kikiri-tts's stage-1/stage-2 recipe updates the full backbone. Text encoder, PL-BERT text-understanding model, prosody/duration predictor, and the ISTFTNet vocoder all receive gradient updates during training, directly exposed to German audio paired with German IPA phoneme sequences (produced by espeak-ng --language=de during training data prep).

The single voice shipped here, df_kerstin, was fine-tuned on Kerstin 1.0 (CC0-1.0), a ~2.3 hour single-speaker German TTS corpus stewarded by the Open Home Foundation and recorded with the Piper Recording Studio tool. df_kerstin follows Kokoro's existing single-letter language-prefix voice-naming convention (d for German, f for female), extended here since upstream Kokoro never shipped a German voice. The kerstin name mirrors the dataset it was fine-tuned on, and coincidentally also matches an unrelated Piper voice name (de_DE-kerstin-low). The two are different model formats with no functional relationship.

Because the fine-tune's training-time phonemes came from espeak-ng's German output rather than Crane's own G2P, differences in IPA conventions (stress placement, diacritic usage, r-realization) between the two may make this model render Crane's G2P output slightly less naturally than it would render espeak-ng's own output for the same text. This is a phoneme-distribution mismatch, not an intelligibility problem. It hasn't produced broken output in testing so far.

Inference follows the same ONNX input/output contract as upstream Kokoro exports (input_ids, style, speed in; audio samples out at 24kHz), so any Kokoro-compatible ONNX runtime should be able to load onnx/model.onnx and voices/df_kerstin.bin. This model is built and tested for Crane, whose GermanG2p engine produces the German IPA input this model expects. It has not been verified against other runtimes, but since Crane's German G2P support originates from the same phonemization assets used for Moonshine ASR, it is also likely to work with Moonshine TTS if paired with a compatible German phonemizer.

Apache 2.0, matching both Kokoro-82M and kikiri-tts. The df_kerstin training data, Kerstin 1.0, is CC0-1.0.